net/core/dst.c

   1 /*
   2  * net/core/dst.c       Protocol independent destination cache.
   3  *
   4  * Authors:             Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  */
   7
   8 #include <linux/bitops.h>
   9 #include <linux/errno.h>
  10 #include <linux/init.h>
  11 #include <linux/kernel.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/mm.h>
  14 #include <linux/module.h>
  15 #include <linux/slab.h>
  16 #include <linux/netdevice.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/string.h>
  19 #include <linux/types.h>
  20 #include <net/net_namespace.h>
  21 #include <linux/sched.h>
  22 #include <linux/prefetch.h>
  23
  24 #include <net/dst.h>
  25
  26 /*
  27  * Theory of operations:
  28  * 1) We use a list, protected by a spinlock, to add
  29  *    new entries from both BH and non-BH context.
  30  * 2) In order to keep spinlock held for a small delay,
  31  *    we use a second list where are stored long lived
  32  *    entries, that are handled by the garbage collect thread
  33  *    fired by a workqueue.
  34  * 3) This list is guarded by a mutex,
  35  *    so that the gc_task and dst_dev_event() can be synchronized.
  36  */
  37
  38 /*
  39  * We want to keep lock & list close together
  40  * to dirty as few cache lines as possible in __dst_free().
  41  * As this is not a very strong hint, we dont force an alignment on SMP.
  42  */
  43 static struct {
  44         spinlock_t              lock;
  45         struct dst_entry        *list;
  46         unsigned long           timer_inc;
  47         unsigned long           timer_expires;
  48 } dst_garbage = {
  49         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  50         .timer_inc = DST_GC_MAX,
  51 };
  52 static void dst_gc_task(struct work_struct *work);
  53 static void ___dst_free(struct dst_entry *dst);
  54
  55 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  56
  57 static DEFINE_MUTEX(dst_gc_mutex);
  58 /*
  59  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  60  */
  61 static struct dst_entry         *dst_busy_list;
  62
  63 static void dst_gc_task(struct work_struct *work)
  64 {
  65         int    delayed = 0;
  66         int    work_performed = 0;
  67         unsigned long expires = ~0L;
  68         struct dst_entry *dst, *next, head;
  69         struct dst_entry *last = &head;
  70
  71         mutex_lock(&dst_gc_mutex);
  72         next = dst_busy_list;
  73
  74 loop:
  75         while ((dst = next) != NULL) {
  76                 next = dst->next;
  77                 prefetch(&next->next);
  78                 cond_resched();
  79                 if (likely(atomic_read(&dst->__refcnt))) {
  80                         last->next = dst;
  81                         last = dst;
  82                         delayed++;
  83                         continue;
  84                 }
  85                 work_performed++;
  86
  87                 dst = dst_destroy(dst);
  88                 if (dst) {
  89                         /* NOHASH and still referenced. Unless it is already
  90                          * on gc list, invalidate it and add to gc list.
  91                          *
  92                          * Note: this is temporary. Actually, NOHASH dst's
  93                          * must be obsoleted when parent is obsoleted.
  94                          * But we do not have state "obsoleted, but
  95                          * referenced by parent", so it is right.
  96                          */
  97                         if (dst->obsolete > 0)
  98                                 continue;
  99
 100                         ___dst_free(dst);
 101                         dst->next = next;
 102                         next = dst;
 103                 }
 104         }
 105
 106         spin_lock_bh(&dst_garbage.lock);
 107         next = dst_garbage.list;
 108         if (next) {
 109                 dst_garbage.list = NULL;
 110                 spin_unlock_bh(&dst_garbage.lock);
 111                 goto loop;
 112         }
 113         last->next = NULL;
 114         dst_busy_list = head.next;
 115         if (!dst_busy_list)
 116                 dst_garbage.timer_inc = DST_GC_MAX;
 117         else {
 118                 /*
 119                  * if we freed less than 1/10 of delayed entries,
 120                  * we can sleep longer.
 121                  */
 122                 if (work_performed <= delayed/10) {
 123                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 124                         if (dst_garbage.timer_expires > DST_GC_MAX)
 125                                 dst_garbage.timer_expires = DST_GC_MAX;
 126                         dst_garbage.timer_inc += DST_GC_INC;
 127                 } else {
 128                         dst_garbage.timer_inc = DST_GC_INC;
 129                         dst_garbage.timer_expires = DST_GC_MIN;
 130                 }
 131                 expires = dst_garbage.timer_expires;
 132                 /*
 133                  * if the next desired timer is more than 4 seconds in the
 134                  * future then round the timer to whole seconds
 135                  */
 136                 if (expires > 4*HZ)
 137                         expires = round_jiffies_relative(expires);
 138                 schedule_delayed_work(&dst_gc_work, expires);
 139         }
 140
 141         spin_unlock_bh(&dst_garbage.lock);
 142         mutex_unlock(&dst_gc_mutex);
 143 }
 144
 145 int dst_discard_sk(struct sock *sk, struct sk_buff *skb)
 146 {
 147         kfree_skb(skb);
 148         return 0;
 149 }
 150 EXPORT_SYMBOL(dst_discard_sk);
 151
 152 const struct dst_metrics dst_default_metrics = {
 153         /* This initializer is needed to force linker to place this variable
 154          * into const section. Otherwise it might end into bss section.
 155          * We really want to avoid false sharing on this variable, and catch
 156          * any writes on it.
 157          */
 158         .refcnt = ATOMIC_INIT(1),
 159 };
 160
 161
 162 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
 163                 int initial_ref, int initial_obsolete, unsigned short flags)
 164 {
 165         struct dst_entry *dst;
 166
 167         if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
 168                 if (ops->gc(ops))
 169                         return NULL;
 170         }
 171         dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
 172         if (!dst)
 173                 return NULL;
 174         dst->child = NULL;
 175         dst->dev = dev;
 176         if (dev)
 177                 dev_hold(dev);
 178         dst->ops = ops;
 179         dst_init_metrics(dst, dst_default_metrics.metrics, true);
 180         dst->expires = 0UL;
 181         dst->path = dst;
 182         dst->from = NULL;
 183 #ifdef CONFIG_XFRM
 184         dst->xfrm = NULL;
 185 #endif
 186         dst->input = dst_discard;
 187         dst->output = dst_discard_sk;
 188         dst->error = 0;
 189         dst->obsolete = initial_obsolete;
 190         dst->header_len = 0;
 191         dst->trailer_len = 0;
 192 #ifdef CONFIG_IP_ROUTE_CLASSID
 193         dst->tclassid = 0;
 194 #endif
 195         atomic_set(&dst->__refcnt, initial_ref);
 196         dst->__use = 0;
 197         dst->lastuse = jiffies;
 198         dst->flags = flags;
 199         dst->pending_confirm = 0;
 200         dst->next = NULL;
 201         if (!(flags & DST_NOCOUNT))
 202                 dst_entries_add(ops, 1);
 203         return dst;
 204 }
 205 EXPORT_SYMBOL(dst_alloc);
 206
 207 static void ___dst_free(struct dst_entry *dst)
 208 {
 209         /* The first case (dev==NULL) is required, when
 210            protocol module is unloaded.
 211          */
 212         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 213                 dst->input = dst_discard;
 214                 dst->output = dst_discard_sk;
 215         }
 216         dst->obsolete = DST_OBSOLETE_DEAD;
 217 }
 218
 219 void __dst_free(struct dst_entry *dst)
 220 {
 221         spin_lock_bh(&dst_garbage.lock);
 222         ___dst_free(dst);
 223         dst->next = dst_garbage.list;
 224         dst_garbage.list = dst;
 225         if (dst_garbage.timer_inc > DST_GC_INC) {
 226                 dst_garbage.timer_inc = DST_GC_INC;
 227                 dst_garbage.timer_expires = DST_GC_MIN;
 228                 mod_delayed_work(system_wq, &dst_gc_work,
 229                                  dst_garbage.timer_expires);
 230         }
 231         spin_unlock_bh(&dst_garbage.lock);
 232 }
 233 EXPORT_SYMBOL(__dst_free);
 234
 235 struct dst_entry *dst_destroy(struct dst_entry * dst)
 236 {
 237         struct dst_entry *child;
 238
 239         smp_rmb();
 240
 241 again:
 242         child = dst->child;
 243
 244         if (!(dst->flags & DST_NOCOUNT))
 245                 dst_entries_add(dst->ops, -1);
 246
 247         if (dst->ops->destroy)
 248                 dst->ops->destroy(dst);
 249         if (dst->dev)
 250                 dev_put(dst->dev);
 251         kmem_cache_free(dst->ops->kmem_cachep, dst);
 252
 253         dst = child;
 254         if (dst) {
 255                 int nohash = dst->flags & DST_NOHASH;
 256
 257                 if (atomic_dec_and_test(&dst->__refcnt)) {
 258                         /* We were real parent of this dst, so kill child. */
 259                         if (nohash)
 260                                 goto again;
 261                 } else {
 262                         /* Child is still referenced, return it for freeing. */
 263                         if (nohash)
 264                                 return dst;
 265                         /* Child is still in his hash table */
 266                 }
 267         }
 268         return NULL;
 269 }
 270 EXPORT_SYMBOL(dst_destroy);
 271
 272 static void dst_destroy_rcu(struct rcu_head *head)
 273 {
 274         struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head);
 275
 276         dst = dst_destroy(dst);
 277         if (dst)
 278                 __dst_free(dst);
 279 }
 280
 281 void dst_release(struct dst_entry *dst)
 282 {
 283         if (dst) {
 284                 int newrefcnt;
 285                 unsigned short nocache = dst->flags & DST_NOCACHE;
 286
 287                 newrefcnt = atomic_dec_return(&dst->__refcnt);
 288                 WARN_ON(newrefcnt < 0);
 289                 if (!newrefcnt && unlikely(nocache))
 290                         call_rcu(&dst->rcu_head, dst_destroy_rcu);
 291         }
 292 }
 293 EXPORT_SYMBOL(dst_release);
 294
 295 static struct kmem_cache *metrics_cache;
 296
 297 void *dst_alloc_metrics(gfp_t flags)
 298 {
 299         return kmem_cache_alloc(metrics_cache, flags);
 300 }
 301 EXPORT_SYMBOL(dst_alloc_metrics);
 302
 303 void dst_free_metrics(void *metrics)
 304 {
 305         kmem_cache_free(metrics_cache, metrics);
 306 }
 307 EXPORT_SYMBOL(dst_free_metrics);
 308
 309 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
 310 {
 311         struct dst_metrics *p = dst_alloc_metrics(GFP_ATOMIC);
 312
 313         if (p) {
 314                 struct dst_metrics *old_p = (struct dst_metrics *)__DST_METRICS_PTR(old);
 315                 unsigned long prev, new;
 316
 317                 atomic_set(&p->refcnt, 1);
 318                 memcpy(p->metrics, old_p->metrics, sizeof(p->metrics));
 319
 320                 new = (unsigned long) p;
 321                 prev = cmpxchg(&dst->_metrics, old, new);
 322
 323                 if (prev != old) {
 324                         dst_free_metrics(p);
 325                         p = (struct dst_metrics *)__DST_METRICS_PTR(prev);
 326                         if (prev & DST_METRICS_READ_ONLY)
 327                                 p = NULL;
 328                 } else if (prev & DST_METRICS_REFCOUNTED) {
 329                         if (atomic_dec_and_test(&old_p->refcnt))
 330                                 dst_free_metrics(old_p);
 331                 }
 332         }
 333         BUILD_BUG_ON(offsetof(struct dst_metrics, metrics) != 0);
 334         return (u32 *)p;
 335 }
 336 EXPORT_SYMBOL(dst_cow_metrics_generic);
 337
 338 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
 339 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
 340 {
 341         unsigned long prev, new;
 342
 343         new = ((unsigned long) &dst_default_metrics) | DST_METRICS_READ_ONLY;
 344         prev = cmpxchg(&dst->_metrics, old, new);
 345         if (prev == old)
 346                 dst_free_metrics(__DST_METRICS_PTR(old));
 347 }
 348 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
 349
 350 /**
 351  * __skb_dst_set_noref - sets skb dst, without a reference
 352  * @skb: buffer
 353  * @dst: dst entry
 354  * @force: if force is set, use noref version even for DST_NOCACHE entries
 355  *
 356  * Sets skb dst, assuming a reference was not taken on dst
 357  * skb_dst_drop() should not dst_release() this dst
 358  */
 359 void __skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst, bool force)
 360 {
 361         WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
 362         /* If dst not in cache, we must take a reference, because
 363          * dst_release() will destroy dst as soon as its refcount becomes zero
 364          */
 365         if (unlikely((dst->flags & DST_NOCACHE) && !force)) {
 366                 dst_hold(dst);
 367                 skb_dst_set(skb, dst);
 368         } else {
 369                 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
 370         }
 371 }
 372 EXPORT_SYMBOL(__skb_dst_set_noref);
 373
 374 /* Dirty hack. We did it in 2.2 (in __dst_free),
 375  * we have _very_ good reasons not to repeat
 376  * this mistake in 2.3, but we have no choice
 377  * now. _It_ _is_ _explicit_ _deliberate_
 378  * _race_ _condition_.
 379  *
 380  * Commented and originally written by Alexey.
 381  */
 382 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 383                        int unregister)
 384 {
 385         if (dst->ops->ifdown)
 386                 dst->ops->ifdown(dst, dev, unregister);
 387
 388         if (dev != dst->dev)
 389                 return;
 390
 391         if (!unregister) {
 392                 dst->input = dst_discard;
 393                 dst->output = dst_discard_sk;
 394         } else {
 395                 dst->dev = dev_net(dst->dev)->loopback_dev;
 396                 dev_hold(dst->dev);
 397                 dev_put(dev);
 398         }
 399 }
 400
 401 static int dst_dev_event(struct notifier_block *this, unsigned long event,
 402                          void *ptr)
 403 {
 404         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 405         struct dst_entry *dst, *last = NULL;
 406
 407         switch (event) {
 408         case NETDEV_UNREGISTER_FINAL:
 409         case NETDEV_DOWN:
 410                 mutex_lock(&dst_gc_mutex);
 411                 for (dst = dst_busy_list; dst; dst = dst->next) {
 412                         last = dst;
 413                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 414                 }
 415
 416                 spin_lock_bh(&dst_garbage.lock);
 417                 dst = dst_garbage.list;
 418                 dst_garbage.list = NULL;
 419                 spin_unlock_bh(&dst_garbage.lock);
 420
 421                 if (last)
 422                         last->next = dst;
 423                 else
 424                         dst_busy_list = dst;
 425                 for (; dst; dst = dst->next)
 426                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 427                 mutex_unlock(&dst_gc_mutex);
 428                 break;
 429         }
 430         return NOTIFY_DONE;
 431 }
 432
 433 static struct notifier_block dst_dev_notifier = {
 434         .notifier_call  = dst_dev_event,
 435         .priority = -10, /* must be called after other network notifiers */
 436 };
 437
 438 void __init dst_init(void)
 439 {
 440         register_netdevice_notifier(&dst_dev_notifier);
 441         metrics_cache = kmem_cache_create("dst_metrics",
 442                                           sizeof(struct dst_metrics),
 443                                           DST_METRICS_ALIGNMENT,
 444                                           SLAB_PANIC, NULL);
 445 }